One of the documents that you pointed to does contain the kernel of your answer:
"The default rendering order for a layer without continuous rasterization is masks, followed by effects, and then transformations; whereas the default rendering order for a continuously rasterized layer is masks, followed by transformations, and then effects."
The issue is that for continuously rasterized layers, effects follow transformations (e.g., Position changes). You have several ways to take more control of the render order. The two that come immediately to mind are (1) precompose the text layer and move the precomposition layer (2) apply the Transform effect and do your position animation within that effect.
It's a bummer there isn't an expression that will make the Effect stick.
I can't pre-compose because I'm using the result of Continuously Rasterize to zoom ridiculously close to the layer - precomposing without breaking down in rez would require a gi-normous Comp and there are a lot of them.
The Transform effect doesn't totally help - though I understand where you were going with it - because it doesn't have control for the Z position (sliding the Z pos for the LAYER comes back with the same no-stick effect. Plus, who wants to move a layer around scrubbing sliders... ha ha.
If the effect only operates on the RGB, one technique that might work is to use the text layer as a track matte for a solid that has the effects applied to it. The downside is that the effect will not be continuously rasterized and will pixellate as you zoom in (though the shape of the text will not).
It is possible to use expressions to keep position controls pinned to a layer when using continuous rasterization, but its implementation and its effectiveness depends a great deal on which effect you are using.